SID Retreat Demo

To gain insight into the limitations of our automatic speaker ID system, we have prepared a test for humans that represents the trials that yielded errors on our best system.

Best System Performance (Switchboard 1)

System EER DCF
(1) SRI GMM 0.90 0.05092
(2) SRI dur. model system 2.60 0.18967
(3) LM (word bigrams) 9.71 0.57516
(4) Kofi's HMM system 1.25 0.06002
(5) Andy's phone-bigrams system 1.74 0.08435
Approximate linear combination of (1)-(5) 0.27 0.00699

Listening Experiment

These are the trials that our best system has gotten wrong. See if you can beat it!

Test Speaker Target Speaker Answer Machine Score (thresh = 0.03)
Test Speech Conv. 1 Conv. 2 Conv. 3 Conv. 4 Conv. 5 Conv. 6 Conv. 7 Conv. 8    
Test Speech Conv. 1 Conv. 2 Conv. 3 Conv. 4 Conv. 5 Conv. 6 Conv. 7 Conv. 8    
Test Speech Conv. 1 Conv. 2 Conv. 3 Conv. 4 Conv. 5 Conv. 6 Conv. 7 Conv. 8    
Test Speech Conv. 1 Conv. 2 Conv. 3 Conv. 4 Conv. 5 Conv. 6 Conv. 7 Conv. 8    
Test Speech Conv. 1 Conv. 2 Conv. 3 Conv. 4 Conv. 5 Conv. 6 Conv. 7 Conv. 8    
Test Speech Conv. 1 Conv. 2 Conv. 3 Conv. 4 Conv. 5 Conv. 6 Conv. 7 Conv. 8    
Test Speech Conv. 1 Conv. 2 Conv. 3 Conv. 4 Conv. 5 Conv. 6 Conv. 7 Conv. 8    
Test Speech Conv. 1 Conv. 2 Conv. 3 Conv. 4 Conv. 5 Conv. 6 Conv. 7 Conv. 8    

Kofi Boakye
Last modified: Wed Jul 21 17:59:43 PDT 2004